Vercel’s Sol Price Cut Came With a New Storage Bill

The GPT-5.6 Sol price cut that landed on August 21, 2026 is OpenAI’s own list-price change, not a Vercel promotion. OpenAI’s model page now reads $4 per million input tokens and $20 per million output tokens, down from $5 and $30, and Reuters reported the same figures the same day. Vercel’s AI Gateway changelog is the attestation of how that cut flows through to a platform customer: the Gateway’s existing 50% Sol discount now applies to the lower list, which puts the Default tier at $2 and $10 through September 18.

That is the headline most coverage will carry. It is not the whole changelog. Inside the same August 21 window Vercel also published a price for Deployment Storage, the build output and function bundles it keeps so a team can roll back: $0.10 per GB-month at the Pro list price, with Hobby teams including up to 10 GB and existing teams continuing on their current pricing. One line makes agent traffic cheaper. The other gives a growing team’s deploy history a priced line of its own.

This post sizes the Sol cut against the $5 / $30 baseline we published at GA, lays out the four prices a Sol token can carry on the same afternoon depending on how you route it, reads the Deployment Storage mechanics in Vercel’s own words, and ends with the two numbers on your own account that decide whether the week was a net saving or a net cost for you.

Start with the source that settles the framing. OpenAI’s GPT-5.6 Sol model page states the new rate, the size of the reduction, and the duration in one passage: $4 per million input tokens and $20 per million output tokens, “a 20% reduction in input pricing and a 33% reduction in output pricing,” with the promotional pricing “available at least through November 21, 2026.” Reuters carried the same numbers on August 21, describing a cut of “more than 20%” for the next three months and listing the previous rate as $5 and $30.

Those previous figures are the ones to anchor on, because they match what we published ourselves. Our GA coverage of Sol’s GA pricing recorded Sol at $5 / $30 per million, identical to the June 26 preview terms and flat against GPT-5.5. The August 21 cut is therefore precisely sizeable from our own baseline: $5 to $4 on input is a 20% reduction, and $30 to $20 on output is a 33% reduction, or one third. Both are OpenAI list prices before any reseller or gateway discount.

Two smaller details travelled with the cut, and they carry different weights of evidence. The same Reuters wire report also recorded cached input tokens dropping from $0.50 to $0.40 per million, a 20% reduction proportional to the input cut. We have not confirmed that cached figure on OpenAI’s own page, so treat it as reported rather than double-sourced. Reuters also framed the cut as competitive, writing that OpenAI “faces growing competition from Anthropic and Chinese AI models.” That is the wire service’s editorial read, not a quoted OpenAI statement of motive, and we present it as such.

For context, that report’s own comparison placed Anthropic’s Claude Opus 5 at $5 input and $25 output per million, and Claude Fable 5 at $10 and $50. On those reported figures Sol’s new list sits below both on input and below both on output. Whether that comparison holds at your effort level and context length is a routing question, not a list-price question, and our cost-optimization routing playbook covers how to test it.

AI Gateway’s base mechanic is pass-through. Vercel’s AI Gateway pricing docs state there is no markup and no platform fee on tokens, so when a provider lowers its list price the Gateway price moves with it automatically. That is why Vercel’s August 21 entry is an attestation rather than an announcement: OpenAI moved the list, and the Gateway followed.

What makes the Vercel side interesting is the layer already sitting on top. On August 17, Vercel announced a 50% discount on Sol through AI Gateway, running through September 18, 2026. That promotion predates OpenAI’s cut and is independent of it. When the list dropped four days later, the August 21 changelog confirmed the two stack: the 50% now applies to the new, lower price.

Vercel published the resulting pay prices by service tier. All figures are per million tokens for requests up to 272K tokens, and the table is Vercel’s own, reproduced here because it is the number a Gateway customer is actually billed, as distinct from the list price the press reports.

Three scope notes from the same entry. The discount applies on every OpenAI service tier, and to cached tokens, cache writes, long-context requests above 272K tokens and US regional rates, all moving by the same proportion. BYOK requests are excluded: if you bring your own OpenAI key through the Gateway, you bill at your own OpenAI rate, which after August 21 is the $4 / $20 list, not the discounted $2 / $10. And the model ID is unchanged at openai/gpt-5.6-sol, so existing integrations pick up both changes with no code edit.

Most write-ups report one before-and-after pair: $5 / $30 became $4 / $20. That is true and incomplete. A Vercel customer riding the existing promotion was already paying $2.50 / $15.00 before the cut, and pays $2.00 / $10.00 after it. A BYOK customer pays the full new list. The table below lays all four out against the original $5 / $30 baseline from our GA post. The percentage columns are our arithmetic on Vercel’s and OpenAI’s stated figures, not a vendor claim.

Read the last row twice. A Gateway customer on the Default tier is paying 40% of the original input price and a third of the original output price, which is $2 of every $5 and $10 of every $30. That is not what the press release says, because the press release is about OpenAI’s list, and it is not what a BYOK customer sees, because the exclusion routes them to list. It is only what an AI Gateway customer on the promotion sees, and only until September 18.

The cut also lands inside a wider August repricing pattern. We tracked the July 30 Luna and Terra cuts and the promo cliffs around them in the August 2026 pricing-cuts tracker, and the cheaper end of the catalog in the cheap-tier repricing wave. Sol is the flagship, and a flagship moving by a third on output is a different signal from a bulk tier moving by 80%: it touches the model that agent products reach for when quality is the constraint, not the one they reach for when volume is.

The second half of the story is in a changelog entry titled “Deployment Storage keeps your deployments rollback-ready,” also dated August 21. Deployment Storage is, per Vercel’s docs, “the build output and Vercel Function bundles retained with your deployments,” which is to say every deployment’s pages, functions and served assets, kept so a team can inspect an earlier deployment or roll back to it. It is a distinct product from Vercel Blob and the other storage lines.

The pricing language is worth quoting rather than paraphrasing, because paraphrases of it have already drifted. Vercel’s text is: Deployment Storage “is billed at $0.10 per GB per month,” “Hobby teams include up to 10 GB,” and “Existing teams continue with their current pricing, with no change to their bill at this time.” The docs add that $0.10 per GB-month is the Pro list price, and that “Enterprise rates follow your contract. Your plan or contract may also include an allowance.”

The unit matters. A GB-month is measured as the maximum stored amount per project on each billing day, summed across the billing period. Vercel’s own worked example: storing 1 GB every day in a 30-day month equals 1 GB-month. Functions Storage is measured the same way and listed at the same Pro price. So the bill tracks what you hold, day by day, not what you upload in a month.

The controls are also documented. Storage volume is governed by the Deployment Retention Policy, which is configurable separately for Pre-Production, Production, Canceled and Errored deployments, and by output size: trimming build output, moving large files to Vercel Blob, and reducing Function bundle size. For a team that ships dozens of preview deployments a day on a large Next.js app, the retention policy for Pre-Production and Canceled deployments is the lever that moves first. Our guide to deploying on Vercel’s AI Cloud covers the build-output side of that.

Neither Vercel’s August 21 changelog nor its Deployment Storage docs state how much a typical team holds, and we are not going to invent a figure. What we can do is show the line item at three stated volumes, held constant for a full 30-day billing month, at Vercel’s stated Pro list price. That matches Vercel’s own “1 GB every day in a 30-day month equals 1 GB-month” example exactly. Read each row as “if your team holds this much, this is the line,” and nothing more.

The arithmetic is deliberately boring. The point is the shape: the line scales linearly with what you retain, and retention is a setting. A team that keeps every preview and canceled deployment indefinitely is choosing the top of that range; a team that prunes Pre-Production retention to a few days is choosing the bottom. The meter does not change what Vercel stores. It changes whether the retention default you never looked at is now a recurring line.

It also bears saying what this is not. It is not Vercel raising prices across the board. Compute on Vercel has moved the other way this year, and Vercel’s other 2026 cost story, Fluid Compute, is the reference point. Deployment Storage is one line item that now carries a published $0.10 per GB-month Pro rate, with existing teams continuing on their current pricing, and that is the full extent of the claim.

Five more entries landed in the same two days, and most of them carry a billing or maturity detail that secondary coverage tends to drop. We list them with Vercel’s own labels, because the labels are where the drift happens.

Two more, briefly. DeepSeek V4 Flash Vision was added to AI Gateway on August 21 under Vercel’s own “Experimental” label, as deepseek/deepseek-v4-flash-vision-exp, with image input and a 1M-token context window; Vercel’s note is that the -exp suffix means behavior may change and a fallback model should stay configured on any production path. We cover the model itself, and its peak and off-peak pricing, in our DeepSeek V4 Flash Vision launch post. And on August 20 the vercel comments CLI shipped, which moves the Toolbar comment loop into the terminal: list and filter unresolved comments scoped to the current branch, inspect a thread, then reply, resolve, reopen, edit or delete.

Everything above reduces to two account-level facts that the headline cannot tell you. The first is which routing path your Sol traffic takes, because that decides whether you are paying $2 or $4 per million input tokens today. The second is how many GB-months of Deployment Storage your projects accrue, because that decides whether the new line is a dollar or a hundred. Both take minutes to read off the dashboard, and both change what you should do next.

One further lever exists for larger spenders. Vercel’s AI Gateway discounts documentation describes custom volume discounts on token spend for six-figure-and-up commitments, varying by model and spend, with ACH billing available. That is separate from and additive to the Sol promotion. If your Sol volume is anywhere near that range, the question is not whether the 50% promo ends on September 18 but what the committed rate looks like after it does. Our AI transformation engagements routinely start with exactly this routing-and-rate review, and the build side, retention included, sits with our web development practice.

Our read of the trend is that the two lines are the same story told from both ends of a platform. Model tokens are a commodity that vendors are competing down: OpenAI cut its flagship by a third on output six weeks after its July 9 GA, Vercel had already discounted the same model by half four days earlier, and a Gateway customer now pays 40% of the July input price. Retained artifacts are the opposite. Every deployment a team keeps is storage the platform holds indefinitely, and indefinite retention is the kind of default that acquires an itemised price once the platform’s own cost of holding it becomes visible. The token side gets cheaper because it is contested; the storage side gets a price because it was never contested at all.

Projecting forward, we would expect the pattern to repeat rather than reverse. The Sol rate is promotional with a November 21 floor, and the Gateway discount ends September 18, so the cheapest price in this post is also the most temporary; a team that rebases its Q4 forecast on $2 / $10 is rebasing on a number with a known expiry. The storage meter, by contrast, has no end date and a phrase, “at this time,” that points the other way. The durable expectation for a platform buyer is that usage-priced compute and tokens keep falling while retained state, whether deploy history, traces or metrics, accumulates small recurring lines. Always-on tracing at $0.50 per million span units and custom metrics billed as observability events, both in the same window, fit that shape exactly.